Freight unmanned aerial vehicle fault diagnosis method based on hybrid driving of multi-relation fault knowledge graph and large model thinking chain prompt

Abstract

The invention discloses a freight unmanned aerial vehicle fault diagnosis method based on hybrid driving of a multi-relation fault knowledge graph and a large model thinking chain prompt, and the method comprises the following steps: constructing a freight unmanned aerial vehicle multi-relation fault knowledge representation, and determining the direct connection between entities and reasoning fault knowledge; entity recognition, relation extraction and knowledge fusion are carried out on the fault knowledge base under the knowledge representation category to complete construction of a multi-relation fault knowledge graph; obtaining a freight unmanned aerial vehicle fault problem, and adopting a large model thinking chain to prompt query; guiding the large model to generate a graph query language through a predefined thinking chain prompt template; and determining the logic relationship of the multi-relationship fault knowledge graph, and carrying out successive reasoning query to obtain an accurate fault reason and a corresponding solution. According to the method, through accurate knowledge modeling and logical reasoning questions and answers, the problems of expression and retrieval of rich relations hidden among fault knowledge are solved, and the strain capacity of a freight unmanned aerial vehicle system to complex factors is improved through rapid and accurate fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of drone fault diagnosis, and more specifically, to a cargo drone fault diagnosis method driven by a hybrid of a multi-relationship fault knowledge graph and a large-model thinking chain prompt. Background Art

[0002] Drone fault diagnosis utilizes various measurement tools to analyze the impact of a UAV system failure in detail. The goal is to identify the fault type, precisely locate the fault, analyze the cause, and provide targeted repair recommendations. Current mainstream methods for UAV fault diagnosis can be broadly categorized into three categories: model-based, signal processing-based, and knowledge-based.

[0003] Knowledge graphs have emerged as an innovative paradigm in the field of knowledge representation. Knowledge graphs are constructed by integrating real-world data, such as fault causes, symptoms, and solutions, into an ontology framework. Graph topology and deductive reasoning based on knowledge graphs facilitate fault diagnosis. Knowledge-based fault diagnosis uses reasoning techniques to provide valuable insights, helping users or applications determine the most appropriate fault resolution. While this reasoning-based approach has not been widely used in drone fault diagnosis, it has demonstrated powerful reasoning capabilities in question-answering, search, and recommendation systems, effectively helping to solve problems and retrieve the most relevant answers.

[0004] Currently, there is a lack of systematic research on knowledge graph construction in the field of cargo drone fault diagnosis. The complex coupling relationships in cargo drone fault data pose challenges to knowledge graph construction. Furthermore, within the niche field of cargo drone fault diagnosis, research on the interaction between large models and knowledge graphs is also scarce. The scarcity of cargo drone fault samples hinders large-scale training, resulting in insufficient inference knowledge and capabilities during multi-path exploration. Traditional prompting methods, when processing multiple knowledge base triples, lack the ability to construct step-by-step reasoning paths and constrain reasoning. Summary of the Invention

[0005] The present invention provides a cargo drone fault diagnosis method that is driven by a hybrid of a multi-relational fault knowledge graph and a large-model thinking chain prompt. It solves the problems of insufficient representation of complex relational knowledge in fault diagnosis by existing binary knowledge graphs, and insufficient reasoning ability of traditional large models for multi-relational fault knowledge graphs in the context of cargo drones.

[0006] To achieve the above objectives, the present invention adopts a cargo drone fault diagnosis method driven by a hybrid of a multi-relationship fault knowledge graph and a large model thinking chain prompt, comprising the following steps:

[0007] Construct a multi-relational fault knowledge representation for cargo drones, including structural and semantic relations to clarify direct connections between entities and reason about fault knowledge;

[0008] Under the scope of knowledge representation, the fault knowledge base is subjected to entity recognition, relationship extraction and knowledge fusion to complete the construction of multi-relationship fault knowledge graph;

[0009] Obtaining cargo drone failure issues uses a large model thinking chain prompt query;

[0010] Use predefined thought chain prompt templates to guide large models to generate graph query language;

[0011] Clarify the logical relationships of the multi-relationship fault knowledge graph and conduct sequential reasoning queries to obtain the precise cause of the fault and the corresponding solution.

[0012] Furthermore, when constructing the multi-relationship fault knowledge representation of cargo drones, the structural relations include the connection between entities and attributes "Has", and the inclusion relationship between entities "Contain"; the semantic relations include the parallel association between fault phenomena and causes "phenomenon-cause", the parallel association between fault causes and solutions "cause-solution", the parallel association between fault phenomena and solutions "phenomenon-solution", and the coupled association between fault phenomena, causes and solutions "multiple phenomena-multiple causes-multiple solutions".

[0013] Furthermore, the BERT-BiLSTM-MHA-CRF deep neural network is used as the basis for entity recognition to build a fault text maintenance data information entity extraction model to represent multi-relational fault entities.

[0014] Furthermore, ERNIE is used to extract fault relations during relation extraction. Through a masking mechanism with prior knowledge, the semantic information of words and phrases is modeled, and entities and phrases are used as masking units for model semantic representation.

[0015] Furthermore, during knowledge fusion, the "CountVectorizer" in the "sklearn" library was used to convert text into vector representation, and the "cosine_similarity" function was used to calculate the cosine similarity between fault descriptions. Similar descriptions were merged into the same category based on the set similarity threshold.

[0016] Furthermore, a large-scale model thinking chain prompt query is used to obtain the cargo drone failure problem. First, a comprehensive analysis is performed on the relationship path of the cargo drone multi-relationship fault knowledge graph, and the relationship between adjacent nodes on the path is expressed as a logical statement. Through relationship path analysis and logical statements, a thinking chain prompt template for step-by-step reasoning and analysis of multiple relationships is set for the large model.

[0017] Furthermore, a predefined thinking chain prompt template is constructed to guide the large model to generate a graph query language. The task of the prompt layer is determined by the identity command, that is, the fault information raised in the user's question is converted into the corresponding Cypher query statement.

[0018] Furthermore, the logical relationship of the multi-relationship fault knowledge graph is clarified to conduct successive reasoning queries. The prompt layer receives the questions input by the user and extracts key elements through semantic analysis, including the fault subject name and related attributes. Based on these key elements and according to the structure and semantic relationship of the knowledge graph, the prompt layer uses the thinking chain prompt template to guide the LLM to generate Cypher query statements that conform to the Neo4j knowledge graph syntax. By reasoning on the graph nodes one by one, multiple causes of the fault and corresponding solutions are obtained.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] The present invention accurately represents complex fault scenarios by constructing multi-relationship chains, and supports hierarchical modeling of multi-factor coupled faults through composite relationship types such as "phenomenon-cause-cause" and "cause-solution-solution". Compared with binary relationships, the comprehensive representation of multiple relationships enables it to more accurately track root causes and generate collaborative solutions in multi-hop reasoning tasks. In addition, the thinking chain prompt technology is introduced to the large model, and a small number of sample prompts are used to decompose the complex multi-variable fault relationship into logically coherent intermediate steps, and construct a step-by-step reasoning path, which effectively solves the problems of insufficient reasoning depth and scarcity of fault samples in large-scale training of traditional LLM models. The present invention lays an important foundation for cargo drones to quickly restore their functions when facing various challenges, improves the feasibility of fault solution methods, and thus significantly improves the resilience of cargo drone systems to complex factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of a flow chart of a hybrid-drive cargo drone fault diagnosis method according to an embodiment of the present invention;

[0022] Figure 2 This is a technical diagram of the integration of a multi-relationship fault knowledge graph and a large model thinking chain prompt according to an embodiment of the present invention;

[0023] Figure 3 A schematic diagram of constructing a multi-relationship fault knowledge graph according to an embodiment of the present invention;

[0024] Figure 4 This is a schematic diagram of the multi-relationship fault knowledge graph thinking chain path construction in an embodiment of the present invention;

[0025] Figure 5This is a schematic diagram of how a predefined thought chain prompt template in an embodiment of the present invention guides a large model to generate a graph query language. DETAILED DESCRIPTION

[0026] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0027] This experimental invention embodiment provides a cargo drone fault diagnosis method driven by a hybrid of multi-relationship fault knowledge graph and large model thinking chain prompts, such as Figure 2 As shown, the following steps are included:

[0028] Step 1: Build a multi-relationship fault knowledge graph based on the cargo drone fault database;

[0029] Step 2: Based on the knowledge features of the multi-relationship fault knowledge graph, the large model is provided with thought chain prompt knowledge to perform fault problem diagnosis.

[0030] Step 1: The specific construction method of multi-relationship fault knowledge graph is as follows: Figure 3 ,It is specifically divided into three steps: multi-relational fault knowledge representation, fault knowledge extraction, and knowledge fusion;

[0031] (1) Multi-relationship fault knowledge representation: The fault knowledge representation dataset not only contains key information such as fault phenomena, causes, and solutions, but also presents specific structural and textual features. In addition, the structural and semantic relationships in the fault information are defined in the cargo drone fault dataset. Structural relationships: Involving the basic connections of "Has" and "Contain", clarifying the direct connection between entities; semantic relationships: More complex and multi-dimensional, including "phenomenon-cause", "cause-solution", "phenomenon-solution", and comprehensive "multiple phenomena-multiple causes-multiple solutions";

[0032] (2) Fault knowledge extraction: Fault knowledge extraction includes entity extraction based on the BERT-BiLSTM-MHA-CRF model and relationship extraction based on the ERNIE model;

[0033] Entity extraction based on the BERT-BiLSTM-MHA-CRF model involves the following steps: When the BERT embedding layer extracts named entities from cargo drone fault knowledge, the BERT model's input layer consists of three parts: word vectors, sentence vectors, and position vectors. The BiLSTM layer, composed of a forward LSTM model and a backward LSTM model, can capture long-term dependencies and contextual information in fault knowledge sequences, effectively encode the contextual information of words in sentences, and effectively learn the temporal features of fault text sequences. The CRF layer decodes the fault fusion feature data output by the multi-head attention mechanism layer, obtains the transition probability of each sequence through a feature function, and calculates the sequence label with the highest score using the Viterbi algorithm.

[0034] Relation extraction based on the ERNIE model uses a masking mechanism with prior knowledge. By modeling semantic information such as words and phrases, entities and phrases are used as masking units to enhance the semantic representation capability of the model.

[0035] (3) Knowledge fusion: The knowledge fusion of cargo drone fault diagnosis is to link the semantic understanding of different identification entities in different data sources to the same entity, and realize the disambiguation and coreference resolution of various entity semantics such as homonyms, multiple names and abbreviations. The cosine similarity algorithm is used to calculate the similarity value between candidate entities. The higher the similarity, the closer the two expressions are. The algorithm uses the "CountVectorizer" in the "sklearn" library to convert text into vector representation, and uses the "cosine_similarity" function to calculate the cosine similarity between fault descriptions. Similar descriptions are merged into the same category according to the set similarity threshold.

[0036] Step 2: Based on the knowledge features of the fault multi-relation knowledge graph, the large model is provided with thought chain prompts to perform fault diagnosis.

[0037] Reference Figure 4 By comprehensively analyzing the relational paths of the cargo drone multi-relational fault knowledge graph, the relationships between adjacent nodes on the path are represented as logical statements. The structure of these statements is as follows: "Fault node 1 is connected to nodes A and a through the relationship "fault cause"; nodes A and a are connected to nodes B and b through the relationship "solution"" and so on;

[0038] Reference Figure 5,construct a predefined thinking chain prompt template to guide the large model to generate a graph query language.,Take the fault "the motor cannot start normally" as an example.,First, locate the fault subject node "CPU", and then retrieve the corresponding fault phenomenon node "the motor cannot start normally" from the problem description.,The analysis process includes two key steps: (1) comprehensively identify all cause nodes related to the fault, (2) retrieve the corresponding solution node for each cause;

[0039] By clarifying the logical relationships within the multi-relational fault knowledge graph and conducting sequential reasoning queries, the prompt layer receives user input and extracts key elements through semantic parsing, including the fault subject name and related attributes. Based on these key elements and the knowledge graph structure and semantic relationships, the prompt layer uses a thought chain prompt template to guide the LLM to generate Cypher query statements that conform to the Neo4j knowledge graph syntax. The "AutoTokenizer" and "AutoModel" converter libraries are used to load pre-trained models and related vocabulary. The "generate_cypher_query" function then converts the user query into a targeted Cypher query. By sequentially reasoning on the graph nodes, the multiple causes of the fault and their corresponding solutions are obtained.

[0040] This invention constructs a cargo drone fault diagnosis method driven by a hybrid of a multi-relational fault knowledge graph and large-model thought chain prompts. The multi-relational fault knowledge graph accurately represents complex fault scenario knowledge, more accurately tracking root causes and generating collaborative solutions. By introducing thought chain prompt templates into the large model, complex multivariate fault relationships are decomposed into logically coherent intermediate steps and a step-by-step reasoning path is constructed, effectively addressing the problems of insufficient reasoning depth and scarce fault samples in large-scale training of traditional LLM models. This invention lays an important foundation for drones to quickly restore their functions in the face of various challenges and improves the feasibility of fault resolution solutions.

[0041] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A cargo drone fault diagnosis method driven by a hybrid of multi-relational fault knowledge graph and large-model thinking chain prompts, characterized by: The following steps are involved: Construct a multi-relational fault knowledge representation for cargo drones, including structural and semantic relations to clarify direct connections between entities and reason about fault knowledge; Under the scope of knowledge representation, the fault knowledge base is subjected to entity recognition, relationship extraction and knowledge fusion to complete the construction of multi-relationship fault knowledge graph; Obtaining cargo drone failure issues uses a large model thinking chain prompt query; Use predefined thought chain prompt templates to guide large models to generate graph query language; Clarify the logical relationships of the multi-relationship fault knowledge graph and conduct sequential reasoning queries to obtain the precise cause of the fault and the corresponding solution.

2. The method for fault diagnosis of a cargo drone driven by a hybrid of a multi-relationship fault knowledge graph and a large model thinking chain prompt as claimed in claim 1, characterized in that: When constructing the multi-relationship fault knowledge representation for cargo drones, structural relations include the "Has" connection between entities and attributes, and the "Contain" relationship between entities. Semantic relations include the parallel association between fault phenomena and causes ("phenomenon-cause"), the parallel association between fault causes and solutions ("cause-solution"), the parallel association between fault phenomena and solutions ("phenomenon-solution"), and the coupled association between fault phenomena, causes, and solutions ("multi-phenomenon-multi-cause-multi-solution").

3. The method for fault diagnosis of a cargo drone driven by a hybrid of a multi-relationship fault knowledge graph and a large model thinking chain prompt as described in claim 2 is characterized in that: In entity recognition, the BERT-BiLSTM-MHA-CRF deep neural network is used as the basis to build a fault text maintenance data information entity extraction model to represent multi-relationship fault entities.

4. The method for fault diagnosis of a cargo drone driven by a hybrid of a multi-relationship fault knowledge graph and a large model thinking chain prompt as described in claim 3 is characterized in that: ERNIE is used to extract fault relations. The semantic information of words and phrases is modeled through a masking mechanism with prior knowledge. Entities and phrases are used as masking units for model semantic representation.

5. The method for fault diagnosis of a cargo drone driven by a hybrid of a multi-relationship fault knowledge graph and a large model thinking chain prompt as described in claim 4 is characterized in that: During knowledge fusion, the "CountVectorizer" in the "sklearn" library was used to convert text into vector representations. The "cosine_similarity" function was used to calculate the cosine similarity between fault descriptions. Similar descriptions were grouped into the same category based on the set similarity threshold.

6. A cargo drone fault diagnosis method driven by a hybrid of a multi-relationship fault knowledge graph and a large model thinking chain prompt as described in claim 5, characterized in that: To obtain the cargo drone fault problem, a large model thinking chain prompt query is used. First, the relationship path of the cargo drone multi-relationship fault knowledge graph is comprehensively analyzed, and the relationship between adjacent nodes on the path is expressed as a logical statement. Through the relationship path analysis and logical statements, a thinking chain prompt template for step-by-step reasoning and analysis of multiple relationships is set for the large model.

7. The method for fault diagnosis of a cargo drone driven by a hybrid of a multi-relationship fault knowledge graph and a large model thinking chain prompt as claimed in claim 6, characterized in that: Construct a predefined thinking chain prompt template to guide the large model to generate a graph query language. The task of the prompt layer is determined by the identity command, that is, converting the fault information raised in the user's question into the corresponding Cypher query statement.

8. The method for fault diagnosis of a cargo drone driven by a hybrid of a multi-relationship fault knowledge graph and a large model thinking chain prompt as claimed in claim 7, characterized in that: Clarify the logical relationships of the multi-relationship fault knowledge graph and conduct sequential reasoning queries. The prompt layer receives questions input by users and extracts key elements through semantic analysis, including the fault subject name and related attributes. Based on these key elements and the structure and semantic relationships of the knowledge graph, the prompt layer uses the thinking chain prompt template to guide the large model to generate Cypher query statements that conform to the Neo4j knowledge graph syntax. By sequentially reasoning on the graph nodes, multiple causes of the fault and corresponding solutions are obtained.

Images